Table of Contents
Sovereign AI on-premise deployment means running AI models on infrastructure your organisation controls, so your data, model weights, hardware and daily operations stay under your authority. For an industrial company, that usually means GPU servers in your own data centre or a national cloud, open-weight models, and security and governance built for operational technology.
Key takeaways
- Sovereignty has four layers: data, model, infrastructure and operations. Physical location alone does not cover them all.
- There are four main deployment options: on-premise, private cloud, air-gapped and sovereign cloud. Each trades control against operating effort.
- GPU memory is the main sizing constraint. A 70-billion-parameter model needs about 140 GB for its weights at bfloat16 precision, according to Hugging Face.
- In Algeria, transfers of personal data abroad need authorisation from the national data protection authority under Law 18-07, as amended by Law 25-11 of 24 July 2025, according to CMS and DLA Piper.
- Start with one or two use cases and a pilot on your own documents before buying hardware at scale.
What does sovereign AI actually mean?
"Sovereign AI" is used loosely. For an industrial buyer, it is more useful to break it into four layers and ask who controls each one.
Data. Where are your documents, sensor histories, formulations and prompts stored and processed? Who can read them, and under which country's law? This includes the less obvious data: the questions your engineers type, the answers the model returns, the logs and the vector index built from your documents.
Model. Do you hold the model weights, or do you call a model you never see? Can you keep using the same version for years, or can the provider change or withdraw it? What does the licence allow?
Infrastructure. Who owns and operates the servers, storage and network? Can the provider be compelled by a foreign authority to hand over data? US law, for example, requires US communications and cloud providers to disclose data in their custody when legally ordered, "regardless of whether such communication, record, or other information is located within or outside of the United States" (18 U.S.C. 2713). This is why a data centre located in your country does not settle the question by itself.
Operations. Who patches the system, updates the model, monitors it and responds to incidents? If the vendor needs permanent remote access to keep it running, part of your sovereignty sits with the vendor.
A deployment is sovereign to the degree that your organisation, or parties under your jurisdiction and contract, controls all four layers. Most real projects make deliberate trade-offs between them. The point is to make those trade-offs knowingly.
Why are industrial companies moving AI on-premise?
Three reasons come up repeatedly in oil and gas, mining, chemistry and manufacturing.
The first is the nature of the data. Drilling programmes, process parameters, product formulations, equipment drawings and incident reports are commercially sensitive, and some are close to critical infrastructure. Many of these documents are exactly what a language model needs to be useful, which is why the question of where they are processed comes first.
The second is regulation. In Algeria, personal data transfers abroad are subject to authorisation, as covered in the regulatory section below. Maintenance logs, HR records and shift reports often contain personal data, even in a plant.
The third is operational: plants run on networks that are segmented, sometimes isolated, and often with limited bandwidth at remote sites. A model that runs locally keeps working when the external link does not.
We cover the strategic case in more depth in why sovereign AI matters for Algerian industry. This guide focuses on how to deploy it.
What are the deployment options for sovereign AI?
There are four main patterns. They are not exclusive; many organisations combine them.
On-premise. GPU servers in your own data centre, managed by your IT team or a contracted integrator. You control the hardware, the network and access. The model can still receive updates over a controlled internet connection.
Private cloud. A dedicated environment, either your own virtualised cluster or a single-tenant deployment at a hosting provider. It offers more elasticity than bare-metal servers. Sovereignty depends on who operates it and under which jurisdiction.
Air-gapped. No network connection to the outside world. Models, software updates and new data arrive through a controlled transfer process. This is the strongest isolation and the heaviest to operate. It is suited to environments near control systems or with highly classified data.
Sovereign cloud. A cloud service operated by a provider under national law, with data centres in the country. In Algeria, a 2017 ARPT decision (Decision 48/SP/PC/ARPT/17) requires cloud hosting providers to "guarantee that customer data is hosted and stored on national territory", according to the CMS expert guide. Sovereign cloud reduces the hardware burden on you, but GPU availability and the provider's operational controls need checking.
How do the options compare?
| Criterion | On-premise | Private cloud | Air-gapped | Sovereign cloud |
|---|---|---|---|---|
| Data stays under your control | Yes | Yes, if single-tenant and you hold the keys | Yes, strongest | Depends on contract and provider |
| Exposure to foreign legal orders | Low | Depends on the operator | Lowest | Low if the provider is national |
| Upfront hardware investment | High | Medium | High | Low |
| Time to first deployment | Weeks to months, including procurement | Weeks | Longest | Shortest, if GPUs are available |
| Model and software updates | Normal, over controlled connection | Normal | Manual, via controlled transfer | Shared with the provider |
| In-house skills needed | GPU operations, MLOps, security | Platform and MLOps | All of the above plus offline update processes | Mainly integration and governance |
| Best fit | Core knowledge assistants, R&D data | Multi-site groups, variable load | Systems near OT, classified data | Pilots, less sensitive workloads |
A common pattern is to run the language model and the document index on-premise for sensitive data, and use a sovereign or private cloud for less sensitive workloads or for burst capacity. For a full cost comparison against cloud APIs, see on-premise LLM vs cloud API: the real cost.
What hardware do you need for an on-premise LLM?
GPU memory is the main constraint. Hugging Face gives a simple rule of thumb: loading a model with X billion parameters needs roughly 2 × X GB of GPU memory at bfloat16 precision, and about double that at float32 (Hugging Face). Its own example: Llama-2-70b needs about 140 GB just for its weights.
Quantization reduces this. In the same guide, a model of about 15.5 billion parameters used around 31 GB in bfloat16, about 15 GB in 8-bit and about 9.5 GB in 4-bit, with little visible loss in output quality in that test, though the guide warns that 4-bit results can differ and inference can be slower (Hugging Face).
Applying the same arithmetic gives a first sizing outline. These figures cover the weights only.
| Model size | Weights at bfloat16 | Weights at 8-bit | Weights at 4-bit |
|---|---|---|---|
| 7 billion parameters | about 14 GB | about 7 GB | about 3.5 GB |
| 24 billion parameters | about 48 GB | about 24 GB | about 12 GB |
| 70 billion parameters | about 140 GB | about 70 GB | about 35 GB |
On top of the weights, you need memory for the key-value cache, which grows with the length of the input and the number of simultaneous users. Hugging Face calculates that, for a 15.5-billion-parameter model with standard multi-head attention, a single 16,000-token input would need about 15 GB of cache, roughly half the size of the weights. Architectures such as multi-query and grouped-query attention reduce this considerably (Hugging Face). Long technical documents and many concurrent users can therefore matter as much as model size.
For reference, NVIDIA lists 80 GB of memory and up to 700 W for the H100 SXM, and 94 GB for the H100 NVL (NVIDIA H100). The L40S has 48 GB and a maximum of 350 W (NVIDIA L40S). Those power figures matter for an industrial site: check rack power, cooling and backup power before ordering, not after.
Do not forget the rest of the stack: fast storage for the document index, enough CPU and RAM for document processing, and a separate environment for testing new model versions before they reach users.
How do you choose the model?
For a sovereign deployment you need open-weight models: models whose weights you can download and run on your own hardware. The main questions are licence, languages and quality on your own documents.
Licence. Licences differ and need a legal read. Mistral Small 3 (24 billion parameters) is published under Apache 2.0, and its model card says it fits on a single RTX 4090 once quantized (Mistral AI on Hugging Face). Qwen2.5-72B-Instruct uses its own "qwen" licence (Qwen on Hugging Face). Llama 3.1 uses a community licence with specific conditions, including a separate licence requirement for services above 700 million monthly active users (Meta).
Languages. Algerian industrial documents mix Arabic, French and English. Check what each model claims, then test it. Qwen2.5-72B-Instruct lists support for more than 29 languages, including French and Arabic (Qwen on Hugging Face). Mistral Small 3's card names French among its languages but does not mention Arabic (Mistral AI on Hugging Face). ALLaM-7B, developed by Saudi Arabia's SDAIA, targets Arabic and English (ALLaM on Hugging Face). Public benchmarks such as the Open Arabic LLM Leaderboard help shortlist models, but they do not measure performance on your maintenance reports or technical vocabulary.
Retrieval before fine-tuning. Most industrial knowledge assistants use retrieval-augmented generation (RAG): the system searches your documents and gives the relevant passages to the model, rather than training the model on them. The approach was described by Lewis et al. in 2020. It keeps your documents in an index you control, makes answers traceable to sources, and lets you update knowledge without retraining. See our guide to RAG for maintenance manuals.
Serving software. Open-source inference servers such as vLLM offer an OpenAI-compatible API and support several quantization formats. An OpenAI-compatible interface makes it easier to replace one model with another later without rewriting your applications.
How do you secure and govern an on-premise AI system?
Running locally removes some risks, such as data leaving for a third-party API, but it adds others. You are now the operator.
Application risks. The OWASP Top 10 for LLM Applications 2025 is a practical checklist. Its first entries are prompt injection, sensitive information disclosure and supply chain risks. It also lists excessive agency and vector and embedding weaknesses, both directly relevant to RAG systems and AI agents.
Controls to put in place:
- Mirror your existing document permissions in the retrieval layer, so a user cannot get through the assistant what they cannot open directly.
- Log queries, retrieved sources and answers, with user identity, and decide who can read those logs.
- Verify the origin and integrity of model files and software before installation, and keep an inventory of versions.
- Give AI agents the narrowest permissions possible, and require human approval for any action that changes a system. Our agentic AI services are built around this principle.
- Keep the AI system out of control loops. If it reads from SCADA or historians, do it through a one-way, read-only path.
Operational technology. If the AI system connects to plant data, apply OT security guidance. NIST SP 800-82 Rev. 3, published in September 2023, covers OT security, including its specific performance, reliability and safety requirements. The ISA/IEC 62443 series sets requirements for securing industrial automation and control systems. Our article on integrating AI with SCADA and OT systems goes into the architecture.
Governance. The NIST AI Risk Management Framework, released on 26 January 2023 for voluntary use, organises AI risk work into four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile on 26 July 2024. Even without a legal obligation to use it, it gives your risk, IT and business teams a shared structure: who owns the system, what it may be used for, how quality is measured and what happens when it fails.
What does Algerian regulation say about AI hosting?
This section summarises published legal guides. It is not legal advice, and the details should be confirmed with counsel against the official texts in the Journal Officiel.
Personal data. Law 18-07 of 10 June 2018 governs the processing of personal data. It was amended and supplemented by Law 25-11 of 24 July 2025 (CMS; DLA Piper). Both guides state that a controller may transfer personal data to a foreign state only with the authorisation of the national authority (ANPDP) and where that state ensures an adequate level of protection, with listed exceptions.
New obligations under Law 25-11. According to the CMS guide, the amendment introduces a data protection officer, impact assessments for high-risk processing, and notification of personal data breaches to the authority within five days (CMS). An AI assistant that indexes HR files or maintenance logs naming employees may fall within these rules.
Cloud hosting. As noted above, the CMS guide reports an ARPT decision requiring cloud hosting providers to keep customer data on national territory. The same guide notes that Algeria has no single cybersecurity law, but a national information systems security framework set by Presidential Decree 20-05.
National AI strategy. Algeria's Scientific Council for Artificial Intelligence announced the adoption of a draft national AI strategy on 8 December 2024, structured around six pillars (Digital Policy Alert). In August 2026, the minister of higher education chaired a meeting launching its implementation, with priorities that include sovereign AI models and securing the processing and storage of public-sector data in protected national data centres (Dzair Tube).
None of this amounts to a general rule that every AI system must run in Algeria. But the direction is clear enough that designing for local hosting from the start is the lower-risk choice for most industrial companies.
What is a realistic step-by-step deployment path?
- Pick one or two use cases with a named owner. A document assistant for maintenance or HSE, or a decision-support tool for R&D, is easier to measure than "AI for the whole company".
- Classify the data. For each use case, list the data sources, who owns them, whether they contain personal data, and whether they may leave the site. This decides your deployment option more than any technical preference.
- Choose the deployment pattern. Use the comparison table above. Many groups start on-premise for sensitive documents and keep the option of a sovereign cloud for other workloads.
- Shortlist and test models on your own documents. Build a test set of 50 to 100 real questions from your engineers, in the languages they actually use, with reference answers. Compare two or three models, quantized and not.
- Size and procure the hardware. Size from the test results: model size, quantization level, document length and expected concurrent users. Plan power, cooling and spare capacity for a test environment.
- Build a pilot with security in place from day one. Access control, logging, a read-only connection to source systems, and an incident contact. Run a short adversarial test using the OWASP list.
- Measure, then decide. Track answer accuracy against your test set, usage, and time saved on real tasks. Decide on scale-up from that evidence, not from the enthusiasm of the pilot team.
- Industrialise operations. Write down who updates models and documents, how changes are tested, how the system is monitored and how it is backed up. In air-gapped sites, define the transfer process for updates.
If you want help with steps 1 to 4, our AI consulting service runs this assessment, and DALILE is our on-premise language model platform for industrial documents.
What are the most common mistakes?
- Treating sovereignty as a location question only. A server in the country operated by a foreign provider, or a vendor with permanent remote access, leaves important layers outside your control.
- Buying GPUs before testing. Hardware bought without a real test set is often the wrong size, too small for long documents or too large for the actual load.
- Sizing for the model and forgetting the users. The key-value cache for long inputs and many concurrent sessions can add gigabytes per long request on top of the weights, depending on the model architecture.
- Not reading the model licence. Licences vary, and some include conditions that legal teams need to approve.
- Testing only in English. A model that performs well on English benchmarks can still struggle with mixed Arabic and French technical text.
- Giving the AI write access to plant systems. Start read-only. Any action that changes a system should require a person to approve it.
- No owner after go-live. Models, documents and access rights change. Without a team responsible for operations, quality degrades quietly.
- Skipping the data protection review. If personal data is involved, an impact assessment may be required under the amended Law 18-07. Check early with your data protection officer or counsel.
What we have seen at KABAS AI
Two published projects show what sovereign deployment looks like in practice. At ADFC, a Sonatrach Group subsidiary, we deployed Virtual Lab to optimise drilling fluid formulations. The deployment keeps data under full control within Sonatrach's IT environment and provides complete traceability of formulation decisions for audit requirements.
At ENAD Shymeca, Virtual Lab reduced physical testing cycles by 40% in detergent formulation R&D, with full data traceability for regulatory compliance. The project was the first Open Innovation contract between an AI startup and a national Algerian industrial company.
In both cases the work started from a specific R&D workflow and the client's own historical data, structured into a reusable repository, rather than from a general AI platform. That is the same order we recommend above: use case and data first, infrastructure second. For sector-specific applications, see our oil and gas solutions.
Sources
- Hugging Face: Optimizing LLMs for speed and memory
- NVIDIA H100 Tensor Core GPU
- NVIDIA L40S GPU
- Mistral Small 3 model card, Hugging Face
- Qwen2.5-72B-Instruct model card, Hugging Face
- ALLaM-7B-Instruct model card, Hugging Face
- Meta Llama 3.1 Community License
- Open Arabic LLM Leaderboard, Hugging Face
- Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, 2020
- vLLM documentation
- OWASP Top 10 for LLM Applications 2025
- NIST SP 800-82 Rev. 3, Guide to Operational Technology Security
- ISA/IEC 62443 series of standards
- NIST AI Risk Management Framework
- 18 U.S.C. 2713, Cornell Legal Information Institute
- CMS Expert Guide to data protection and cyber security laws: Algeria
- DLA Piper Data Protection Laws of the World: Algeria
- Digital Policy Alert: Algerian National Artificial Intelligence Strategy
- Dzair Tube: Kamel Baddari launches implementation of Algeria's national AI strategy
Frequently asked questions
What is the difference between on-premise AI and sovereign AI?
On-premise describes where the hardware sits: in your own facility. Sovereign AI is broader and asks who controls the data, the model, the infrastructure and the day-to-day operation. An on-premise system can still fall short of sovereignty if, for example, the vendor needs remote access to run it or the model licence restricts how you use it.
How much GPU memory do I need to run a large language model on-premise?
Hugging Face's rule of thumb is roughly 2 GB of GPU memory per billion parameters at bfloat16 precision, so a 70-billion-parameter model needs about 140 GB just for its weights. Quantization to 8-bit or 4-bit reduces this substantially. You also need headroom for the key-value cache, which grows with document length and the number of concurrent users.
Is an air-gapped AI deployment realistic for an industrial company?
Yes, for workloads that justify it, such as anything close to control systems or highly confidential process data. Open-weight models and serving software can run with no internet connection. The cost is operational: model updates, security patches and new documents must be brought in through a controlled transfer process that someone owns.
Does Algerian law require AI systems to be hosted in Algeria?
There is no general AI hosting law that we are aware of. Law 18-07 on personal data, as amended by Law 25-11 of 24 July 2025, makes transfers of personal data abroad subject to authorisation by the national data protection authority, and a 2017 ARPT decision requires cloud hosting providers to keep customer data on national territory, according to law-firm summaries. Check your specific case with legal counsel.
Can we start with a cloud API and move on-premise later?
You can, as long as you only send data that is cleared for external processing and you design for portability from the start. Using a serving layer with an OpenAI-compatible API, such as vLLM, makes it easier to swap a cloud endpoint for a local one later. The harder part to move is usually the data pipeline and access control, not the model.



